Low power AI chip for robotics

Researchers in the United States have recently unveiled an innovative low-power multicore AI chip designed specifically for robotics applications. Known as the MAVERIC chip, this cutting-edge technology was developed at the University of California, Berkeley. The chip boasts a 16nm process and features four RISC-V processor cores along with 13 AI accelerator cores. These components work in tandem to deliver impressive AI vision analysis capabilities, achieving a remarkable 72 frames per second for machine learning and robotics tasks.

One of the key strengths of the MAVERIC chip lies in its ability to support 3D reconstruction in robotics applications. By combining depth estimation (DE) and simultaneous localization and mapping (SLAM), the chip enhances perception tasks, paving the way for more advanced and efficient robotic systems.

Presented at the VLSI Technology & Circuits Symposium in Tokyo, the MAVERIC chip can operate at speeds of up to 1 GHz while delivering a peak energy efficiency of 8 TOPS/W. Notably, the chip supports loop closure and is capable of achieving 10 mJ per frame and 72 frames per second during end-to-end DE and SLAM processes.

• DeepX, a prominent player in the field, has announced plans to develop a 5W 2nm edge AI chip, further underscoring the rapid advancements in AI hardware technology.

Equipped with eight 8-bit integer accelerators and five 32-bit floating-point accelerators, the MAVERIC chip also features 2504 Kbits of memory. By offering a supply voltage range from 0.52V to 1.0V, the chip effectively reduces power consumption to as low as 20mW at 100MHz, with power usage peaking at 1.8W when operating at 1GHz.

The architecture of the MAVERIC chip showcases its sophisticated design, highlighting the eight INT8 ML accelerators, five FP32 linear algebra accelerators, four RISC-V CPUs, and a 3-layer NoC that binds these sub-blocks together seamlessly. This intricate configuration enables the chip to deliver exceptional performance and efficiency in AI-driven robotics applications, setting a new standard for intelligent computing in the field.